What is a Chatbot? Definition, Types, and Use Cases (2026)

What a chatbot actually is

A chatbot is a program that converses with users in natural language, replacing or supplementing a human operator. It receives an input (usually text, sometimes voice), interprets what the user wants, and generates a response. The response can be a fixed reply chosen from a flow chart, a database lookup, a multi-step transaction (booking, payment, account lookup), or freely composed text generated by an AI model.

Three things distinguish a chatbot from neighboring categories:

• A chatbot is conversational. A search engine takes a query and returns documents — there is no back-and-forth. A chatbot exchanges turns of message with the user, retaining context across turns.
• A chatbot is automated. Live chat is human-to-human — an agent on the other end. A chatbot replies algorithmically. The two often coexist (a chatbot handles routine questions and hands complex ones off to a live agent), but the chatbot itself is not a human.
• A chatbot is a system, not a single algorithm. It includes a message interface (channel), an interpretation layer (rules or AI), a logic layer (what to say next), and usually integrations with backend systems (CRM, ecommerce, ticketing).

A brief history (why this matters)

The word "chatbot" comes from "chat robot." The first program widely recognized as one — ELIZA, built at MIT by Joseph Weizenbaum in 1966 — used simple pattern matching to imitate a psychotherapist. ELIZA had no understanding of language; it spotted keywords and rephrased the user's sentence as a question. Users nevertheless found it convincing, an effect that came to be called the "ELIZA effect."

For the next several decades chatbots remained niche — IRC bots, customer-service IVR systems, and academic experiments. The modern era of chatbots opened in 2016 when Facebook (now Meta) launched the Messenger Platform, allowing third-party developers to build automated conversations inside Messenger at consumer scale. Manychat, Chatfuel, and a wave of SMB-focused chatbot platforms followed, treating chatbots as marketing automation tools rather than research curiosities.

The current generation, post-2022, is defined by large language models — primarily OpenAI's GPT family, Anthropic's Claude, Google's Gemini, and a growing list of open-weights models. LLMs collapsed two decades of incremental NLP progress into a single broadly capable system, and "chatbot" now almost always implies some LLM behind the scenes — even when the user-facing surface still looks like a button-and-flow interface.

Types of chatbots

There are four primary architectures, and most production systems blend them.

flowchart TB
 subgraph T1[1. Rule-based - decision tree]
 R1[User input] --> R2{Match keyword<br/>or button?}
 R2 -->|Yes| R3[Run scripted reply]
 R2 -->|No| R4[Fallback or escalate]
 end
 subgraph T2[2. NLU-driven - intent classifier]
 N1[User input] --> N2[Classify intent<br/>and extract entities]
 N2 --> N3[Lookup flow for intent]
 N3 --> N4[Run template reply]
 end
 subgraph T3[3. Generative - LLM]
 G1[User input] --> G2[Prompt LLM<br/>with system prompt]
 G2 --> G3[Stream generated reply]
 end
 subgraph T4[4. Hybrid - production default]
 H1[User input] --> H2{Match known<br/>transactional path?}
 H2 -->|Yes| H3[Rule-based flow]
 H2 -->|No| H4[LLM + RAG fallback]
 H3 --> H5[Reply]
 H4 --> H5
 end

Figure 1. Four chatbot architectures. Most 2026 production deployments use the hybrid pattern — predictable flows for known paths (checkout, lead capture), LLM-driven handling for open-ended queries.

1. Rule-based (decision-tree) chatbots

The oldest and still most common type. The chatbot follows a fixed flowchart — if the user says X, reply Y; if the user clicks Button A, go to Step 2. There is no understanding of language; the bot recognizes specific keywords, button taps, or quick-reply selections.

Strengths: Predictable, cheap to run, easy to audit, no hallucination. Excellent for transactional flows (lead capture, booking, ecommerce checkout, FAQ menus).

Weaknesses: Brittle when users deviate from the script, cannot handle open-ended questions, scales poorly as the conversation tree grows.

Most marketing-focused chatbot builders (Manychat, Chatfuel, Landbot) lean heavily on rule-based flows because marketing conversions follow predictable paths.

2. AI-powered (NLU + intent classification) chatbots

These use a Natural Language Understanding model trained on labeled examples to map user messages to a small set of intents ("cancel_order", "check_balance", "book_appointment"), then run a rule-based response for each intent. Google's Dialogflow, Microsoft's Bot Framework, and most enterprise platforms in the 2017-2022 era worked this way.

Strengths: Tolerant of phrasing variation, can handle some open-ended language, easier to audit than a generative chatbot.

Weaknesses: Requires labeled training data, doesn't handle entirely novel questions, falls off a cliff outside trained intents.

3. Generative (LLM-powered) chatbots

A large language model is the response engine — usually wrapped with a system prompt that defines tone, scope, and guardrails. Recent chatbots from Botpress, Chatbase, Intercom Fin, and most Manychat AI features use this architecture.

Strengths: Handles arbitrary phrasing, generates fluent and contextually appropriate replies, can summarize, translate, and reason. No training data labeling required.

Weaknesses: Can hallucinate (generate plausible-sounding but wrong information), more expensive per turn, harder to audit. Often paired with retrieval-augmented generation (RAG) to ground responses in a verified knowledge base.

4. Hybrid chatbots

The dominant production architecture in 2026. The bot uses rule-based flows for known transactional paths (so the checkout flow doesn't hallucinate), NLU for intent classification on free-text queries, and an LLM (often with RAG) for open-ended answers. Most modern SMB platforms offer all three modes inside the same flow builder.

When to use a chatbot

A chatbot is the right tool when:

• High volume of repetitive questions. The 80% of customer questions that follow the same patterns (order status, business hours, return policy, basic product info) can be deflected to a chatbot at near-zero marginal cost.
• 24/7 coverage matters. A chatbot answers at 3 AM the same as at 3 PM. Critical for global businesses and SMB owners who can't staff round-the-clock.
• The conversation is structured. Lead capture, appointment booking, order placement, account lookup — these have a defined goal and a small set of inputs. Chatbots execute these reliably.
• You want to scale messenger marketing. Instagram comment automation, WhatsApp campaign flows, Messenger sequences — chatbots make outreach at scale possible without buying ads.
• Your audience lives on a messaging app. In Brazil, Mexico, and most of South Asia, WhatsApp is the dominant communication channel. A chatbot is the only way to engage that audience automatically.

When NOT to use a chatbot

A chatbot is the wrong tool — or is at minimum insufficient on its own — when:

• The user is upset or the stakes are high. Account cancellations, billing disputes, refund requests, complaints about a defective product — these need a human empathetic response and the authority to make exceptions. A chatbot here makes things worse.
• The conversation is highly novel or technical. A chatbot is good at the common 80%; the unusual 20% needs a human or a senior support specialist.
• Compliance requires a human in the loop. Medical advice, legal advice, regulated financial transactions in many jurisdictions, or anything that requires verified identity confirmation.
• You don't have content to train it on. A chatbot is only as good as its knowledge base. If your product documentation is thin or outdated, the chatbot will sound thin and outdated.

Examples in practice

Instagram lead capture (Manychat). A creator posts a Reel and asks viewers to comment a keyword. Manychat detects the comment, sends an automated DM with a discount link, captures the user's email through a form inside the DM, and adds them to a tagged segment. Total time from comment to lead: under 30 seconds. No human involvement until the prospect is ready to buy.

WhatsApp commerce flow (Wati, AiSensy, Manychat Pro). A returning customer messages a brand's WhatsApp Business number. The chatbot greets them by name (pulled from CRM), offers product categories, sends product cards with images and prices, accepts a purchase intent, and either completes the transaction inside WhatsApp or hands off to a checkout link. Operators in Brazil and India routinely run 60-80% of ecommerce volume through this pattern.

Customer-support deflection (Intercom Fin, Chatbase). A user asks a technical question on a website widget. The chatbot retrieves relevant documentation pages, summarizes them in the user's words, and either resolves the question or escalates to a human agent with the user's full context attached. Deflection rates of 40-65% on common-question categories are typical in 2026.

Related terms

• Conversational AI — the broader umbrella that includes chatbots plus voice assistants, intelligent virtual agents, and other conversational interfaces.
• AI agent — a chatbot that can take actions in external systems (not just reply), increasingly the dominant frontier in 2026.
• Natural Language Processing — the technical field underlying chatbots' ability to interpret language.
• Large language model — the type of AI model powering most modern generative chatbots.
• Retrieval-augmented generation — the technique used to ground LLM chatbots in a verified knowledge base.
• Live chat — the human-staffed alternative to chatbots, often used in tandem.

FAQ

Is a chatbot the same as AI?

Not quite. A chatbot is a software interface for conversation; AI is the technique that can power it. Rule-based chatbots from 1995 had no AI inside. Modern generative chatbots use large language models, which are a form of AI. The chatbot is the product; AI is one (now common) ingredient.

Are chatbots safe to use for customer data?

That depends entirely on the platform's compliance posture, not the technology itself. Platforms that handle EU customer data must comply with GDPR; healthcare data in the US requires HIPAA-compliant infrastructure and signed Business Associate Agreements. Most general-purpose SMB chatbot platforms (Manychat, Chatfuel, SendPulse, Wati) are not HIPAA-compliant. Always check the vendor's trust/compliance page before processing regulated data through their chatbot.

Can a chatbot replace a customer-service team?

A chatbot can handle a meaningful share of routine volume (typically 30-65% in well-tuned deployments) but rarely the entire load. The economic case is augmentation: deflect routine questions, escalate complex or emotional ones to humans, and free up agent capacity for the work that needs human judgment.

Do chatbots work in languages other than English?

Modern LLM-powered chatbots handle dozens of languages competently, though performance varies. Underlying GPT-4 and Claude work well across major languages (Spanish, Portuguese, French, German, Mandarin, Hindi); coverage is thinner for low-resource languages. Platform-level localization (admin UI translation, help docs) is usually narrower than NLU coverage — see specific platform reviews for language-by-language detail.

How much does a chatbot cost?

For SMBs in 2026, monthly-billed entry-tier pricing across our verified market dataset ranges from $9-69/month, depending on platform and feature scope. WhatsApp Business API conversation fees pass through from Meta separately (typically $0.005-0.09 per conversation). Voice channels are usually pay-per-minute ($0.05-0.15/min). Free tiers exist on most major platforms but with material limits (contact caps, channel restrictions). Cross-platform comparison data is in our reviews section.

Sources

• Weizenbaum, Joseph. ELIZA—a computer program for the study of natural language communication between man and machine. Communications of the ACM, January 1966.
• Meta. Messenger Platform. developers.facebook.com/docs/messenger-platform (verified 26 May 2026).
• Anthropic. Model Context Protocol specification. modelcontextprotocol.io (verified 26 May 2026).
• Industry reviews and platform documentation referenced in linked Chatbotscape platform reviews.